Employing the Centering Theory in Pronoun Resolution from the Semantic Perspective
KONG Fang ZHOU GuoDong * ZHU Qiaoming
JiangSu Provincial Key Lab for Computer Information Processing Technology
School of Computer Science and Technology
Soochow University. Suzhou, China 215006
Email: {kongfang, gdzhou, qmzhu}@suda.edu.cn
Abstract
In this paper, we employ the centering theory in pronoun resolution from the semantic perspective. First, diverse semantic role
features with regard to different predicates
in a sentence are explored. Moreover, given
a pronominal anaphor, its relative ranking
among all the pronouns in a sentence, according to relevant semantic role information and its surface position, is incorporated.
In particular, the use of both the semantic
role features and the relative pronominal
ranking feature in pronoun resolution is
guided by extending the centering theory
from the grammatical level to the semantic
level in tracking the local discourse focus.
Finally, detailed pronominal subcategory
features are incorporated to enhance the
discriminative power of both the semantic
role features and the relative pronominal
ranking feature. Experimental results on the
ACE 2003 corpus show that the centeringmotivated features contribute much to pronoun resolution.
1
Introduction
Coreference accounts for cohesion in a text and
is, in a sense, the hyperlink for a natural language. Especially, a coreference instance denotes an identity of reference and holds between
two referring expressions, which can be named
entities, definite noun phrases, pronouns and so
on. Coreference resolution is the process of linking together multiple referring expressions of a
given entity in the world. The key in coreference
resolution is to determine the antecedent for
each referring expression in a text. The ability of
linking referring expressions both within a sentence and across the sentences in a text is critical
*
to discourse and language understanding in general. For example, coreference resolution is a
key task in information extraction, machine
translation, text summarization, and question
answering.
There is a long tradition of research on
coreference resolution within computational linguistics. While earlier knowledge-lean approaches heavily depend on domain and
linguistic knowledge (Carter 1987; Carbonell
and Brown 1988) and have significantly influenced the research, the later approaches usually
rely on diverse lexical, syntactic and semantic
properties of referring expressions (Soon et al.,
2001;Ng and Cardie, 2002; Zhou et al., 2004).
Current research has been focusing on exploiting
semantic information in coreference resolution.
For example, Yang et al (2005) proposed a template-based statistical approach to compute the
semantic compatibility between a pronominal
anaphor and an antecedent candidate, and Yang
and Su (2007) explored semantic relatedness
information from automatically discovered patterns, while Ng (2007) automatically induced
semantic class knowledge from a treebank and
explored its application in coreference resolution.
Particularly, this paper focuses on the centering theory (Sidner,1981;Grosz et al.,1995;
Tetreault,2001), which reveals the significant
impact of the local focus on referring expressions in that the antecedent of a referring expression usually depends on the center of attention
throughout the local discourse segment (Mitkov,1998). Although the centering theory has
been considered as a critical theory and the driving force behind the coreferential phenomena
since its proposal, its application in coreference
resolution (in particular pronoun resolution) has
been somewhat disappointing: it fails to improve
or even harms the performance of the state-of-
Corresponding author
987
Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 987–996,
c
Singapore, 6-7 August 2009. 2009
ACL and AFNLP
the-art coreference resolution systems in previous research (e.g. Yang et al. 2004). This may be
due to that centering was originally proposed as
a model of discourse coherence instead of
coreference.
The purpose of this paper is to employ the
centering theory in pronoun resolution by extending it from the grammatical level to the semantic level. The intuition behind our approach
is that, via determining the semantic roles of
referring expressions in a sentence, such as
agent and patient, we can derive various centering theory-motivated features in tracking the
continuity or shift of the local discourse focus,
thus allowing us to include document-level
event descriptive information in resolving the
coreferential relations between referring expressions.
To the best of our knowledge, this is the first
research, which successfully applies the centering theory in pronoun resolution from the semantic perspective.
The rest of this paper is organized as follows.
Section 2 briefly describes related work in employing the centering theory and semantic information in coreference resolution. Then, the
centering theory is introduced in Section 3 while
Section 4 details how to employ the centering
theory from the semantic perspective. Section 5
reports and discusses the experimental results.
Finally, we conclude our work in Section 6.
2
Related Work
This section briefly overviews the related work
in coreference resolution from both the centering
theory and semantic perspectives.
2.1
In applications of the centering theory to
coreference resolution, representative work includes Brennan et al. (1987), Strube (1998),
Tetreault (1999) and Yang et al. (2004). Brennan
et al. (1987) presented a centering theory-based
formalism in modeling the local focus structure
in discourse and used it to track the discourse
context in binding occurring pronouns to corresponding entities. In particular, a BFP (Brennan,
Friedman and Pollard) algorithm is proposed to
extend the original centering model to include
two additional transitions called smooth shift
and rough shift. Strube (1998) proposed an S-list
model, assuming that a referring expression prefers a hearer-old discourse entity to other hearernew candidates. Tetreault (1999) further advanced the BFP algorithm by adopting a left-toright breadth first walk of the syntactic parse
trees to rank the antecedent candidates. However,
the above methods have not been systematically
evaluated on large annotated corpora, such as
MUC and ACE. Thus their effects are still unclear in real coreference resolution tasks. Yang
et al (2004) presented a learning-based approach
by incorporating several S-list model-based features to improve the performance in pronoun
resolution. It shows that, although including Slist model-based features can slightly boost the
performance in the ideal case (i.e. given the correct antecedents of anaphor’s candidates), it deteriorates the overall performance in F-measure
with slightly higher precision but much lower
recall, in real cases, where the antecedents of
anaphor’s candidates are determined automatically by a separate coreference resolution module.
2.2
Centering Theory
In the literature, there has been much research in
the centering theory and its application to
coreference resolution.
In the centering theory itself, since the original work of Sidner (1979) on immediate focusing of pronouns and the subsequent work of
Joshi and Weinstein (1981) on centering and
inferences, much research has been done, including centering and linguistic realizations
(Cote 1993; Prince and Walker 1995), empirical
and psycholinguistic evaluation of centering
predictions (Gordon et al 1993,1995; Brennan
1995; Walker et al 1998; Kibble 2001), and the
cross-linguistic work on centering (Ziv and
Crosz1994).
988
Semantic Information
It is well known that semantic information plays
a critical role in coreference resolution. Besides
the common practice of employing a thesaurus
(e.g. WordNet) in semantic consistency checking, much research has been done to explore
various kinds of semantic information, such as
semantic similarity (Harabagiu et al 2000), semantic compatibility (Yang et al 2005, 2007),
and semantic class information (Soon et al 2001;
Ng 2007). Although these methods have been
proven useful in coreference resolution, their
contributions are much limited. For example, Ng
(2007) showed that semantic similarity information and semantic agreement information could
only improve the performance of coreference
resolution by 0.6 and 0.5 in F-measure respectively, on the ACE 2003 NWIRE corpus.
3
Centering Theory
The centering theory is a theory about the local
discourse structure that models the interaction of
referential continuity and the salience of discourse entities in the internal organization of a
text. In natural languages, a given entity may be
referred by different expressions and act as different grammatical roles throughout a text. For
example, people often use pronouns to refer to
the main subject of the discourse in focus, which
can change over different portions of the discourse. One main goal of the centering theory is
to track the focus entities throughout a text.
The main claims of the centering theory can
be formalized in terms of Cb (the backwardlooking center), Cf (a list of forward-looking
centers for each utterance Un) and Cp (the preferred center, i.e. the most salient candidate for
subsequent utterances). Given following two
sentences: 1) Susani gave Betsyj a pet hamsterk;
2) Shei reminded herj that such hamstersk were
quite shy. We can have Ub, Uf and Up as follows:
Ub= “Susan”; Uf={“Susan”, “Betsy”, “a pet
hamster”}; Up= “Susan”.
Cb(Un)=Cb(Un-1)
C (U )≠Cb(Un-1)
or Cb(Un-1) undefined b n
Cb(Un)=Cp(Un)
Continue
Smooth Shift
Cb(Un)≠Cp(Un)
Retain
Rough Shift
Table 1: Transitions in the centering theory
Constraints
C1. There is precisely one Cb.
C2. Every element of Cf(Un) must be realized in Un.
C3. Cb(Un) is the highest-ranked element of Cf(Un-1)
that is realized in Un.
Rules
R1. If some element of Cf(Un-1) is realized as a pronoun in Un, then so is Cb(Un).
R2.Transitions have the descending preference order
of “Continue > Retain > Smooth Shift > Rough
Shift”.
Table 2: Constraints and rules in the centering theory
sentence planning, while cohesion, which orders
propositions in a text to maintain referential continuity, is an important matter for text planning.
Finally, the centering theory imposes several
constraints and rules over Cb/Cf and above transitions, as shown in Table 2.
Given the centering theory as described above,
we can draw the following conclusions:
1) The centering theory is discourse-related and
centers are discourse constructs.
2) The backward-looking center Cb of Un depends only on the expressions that constitute
the utterance. That is, it is independent of its
surface position and grammatical roles.
Moreover, it is not constrained by any previous utterance in the segment. While the elements of Cf(Un) are partially ordered to
reflect relative prominence in Un, grammatical role information is often a major determinant in ranking Cf, e.g. in the descending
priority order of “Subject > Object > Others”
in English (Grosz and Joshi, 2001).
3) Psychological research (Gordon et al. 1993)
and cross-linguistic research (Kameyama
1986, 1988; Walker et al. 1990,1994) have
validated that Cb is preferentially realized by
a pronoun in English.
4) Frequent rough shifts would lead to a lack of
local cohesion. To keep local cohesion, people tend to plan ahead and minimize the
number of focus shifts.
In this paper, we extend the centering theory
from the grammatical level to the semantic level
in attempt to better model the continuity or shift
in the local discourse focus and improve the performance of pronoun resolution via centeringmotivated semantic role features.
4
Furthermore, several kinds of focus transitions are defined in terms of two tests: whether
Cb stays the same (i.e. Cb(Un+1)=Cb(Un)), and
whether Cb is realized as the most prominent
referring expression (i.e. Cb(Un=Cp(Un)). We
refer to the first test as cohesion, and the second
test as salience. Therefore, there are four possible combinations, which are displayed in Table
1 and can result in four kinds of transitions,
namely Continue, Retain, Smooth Shift, and
Rough Shift. Obviously, salience, which chooses
a proper verb form to make Cb prominent within
a clause or sentence, is an important matter for
989
Employing Centering Theory from
Semantic Perspective
In this section, we discuss how to employ the
centering theory in pronoun resolution from the
semantic perspective. In Subsection 4.1, we introduce the semantic roles. In Subsection 4.2, we
introduce how to employ the centering theory in
pronoun resolution via semantic role features.
Finally we compare our method with the previous work in Subsection 4.3.
4.1
Semantic Role
A semantic role is the underlying relationship
that a participant has with a given predicate in a
clause, i.e. the actual role a participant plays in
an event, apart from linguistic encoding of the
situation. If, in some situation, someone named
“John” purposely hits someone named “Bill”,
then “John” is the agent and “Bill” is the patient
of the hitting event. Therefore, given the predicate “hit” in both of the following sentences,
“John” has the same semantic role of agent and
“Bill” has the same semantic role of patient: 1)
John hit Bill. 2) Bill was hit by John.
In the literature, labeling of such semantic
roles has been well defined by the SRL (Semantic Role Labeling) task, which first identifies the
arguments of a given predicate and then assigns
them appropriate semantic roles. During the last
few years, there has been growing interest in
SRL. For example, CoNLL 2004 and 2005 have
made this problem a well-known shared task.
However, there is still little consensus in the linguistic and NLP communities about what set of
semantic role labels are most appropriate. Typical semantic roles include core roles, such as
agent, patient, instrument, and adjunct roles
(such as locative, temporal, manner, and cause).
For core roles, only agent and patient are consistently defined across different predicates, e.g. in
the popular PropBank (Palmer et al. 2005) and
the derived version evaluated in the CoNLL
2004 and 2005 shared tasks, as ARG0 and
ARG1.
In this paper, we extend the centering theory
from the grammatical level to the semantic level
for its better application in pronoun resolution
via proper semantic role features due to three
reasons:
Sentence
Grammatical Role Semantic Role
Bob opened the
Bob:
Bob:
door with a key.
SUBJECT
AGENT
The key opened
The key:
The key :
the door.
SUBJECT
INSTRUMENT
The door opened.
The door:
The door:
SUBJECT
PATIENT
Table 3: Relationship between grammatical roles and
semantic roles: an example
1) Semantic roles are conceptual notions,
whereas grammatical roles are morphsyntactic. While the original centering theory
mainly builds from the grammatical perspective and grammatical roles do not always correspond directly to semantic roles (Table 3
shows an example of various semantic roles
which a subject can play), there is a close relationship between semantic roles and grammatical roles. The statistics in the CoNLL
2004 and 2005 shared tasks (Shen and Lapata,
2007) shows that the semantic roles of
990
ARG0/agent and ARG1/patient account for
85% of all arguments and most likely act as
the centers of the local focus structure in discourse due to the close relationship between
subject/object and agent/patient. Therefore, it
is appropriate to model the centers of an utterance from the semantic perspective via
semantic roles.
2) In a sense, semantic roles imply the information of grammatical roles, especially for subject/object. For example, the position of an
argument and the voice of the predicate verb
play a central role in SRL. In intuition, an argument, which occurs before an active verb
and has the semantic role of Arg0/agent,
tends to be a subject. That is to say, semantic
roles (e.g. Arg0/agent and Arg1/patient) can
be mapped into their corresponding grammatical roles (e.g. subject and object), using
some heuristic rules. Therefore, it would be
interesting to represent the centers of the utterances and employ the centering theory
from the semantic perspective.
3) Semantic role labeling has been well studied
in the literature and there are good ready-touse toolkits available. For example, Pradhan
(2005) achieved 82.2 in F-measure on the
CoNLL 2005 version of the Propbank. In
contrast, the research on grammatical role labeling is much less with the much lower
state-of-the-art performance of 71.2 in Fmeasure (Buchholz, 1999). Therefore, it may
be better to explore the centering theory from
the semantic perspective.
4.2
Designing Centering-motivated
tures from Semantic Perspective
Fea-
In this paper, the centering theory is employed in
pronoun resolution via three kinds of centeringmotivated features:
1) Semantic role features. They are achieved by
checking possible semantic roles of referring
expressions with regard to various predicates
in a sentence. Due to the close relationship
between subject/object and agent/patient, semantic role information should be also a major determinant in deciding the center of an
utterance, which is likely to be the antecedent
of a referring expression in the descending
priority order of “Agent > Patient > Others”
with regard to their semantic roles, corresponding to the descending priority order of
“Subject > Object > Others” with regard to
their grammatical roles.
2) Relative pronominal ranking feature. Due to
the predominance of pronouns in tracking the
local discourse structure 1 , the relative ranking of a pronoun among all the pronouns in a
sentence should be useful in pronoun resolution. This is realized in this paper according
to its semantic roles (with regard to various
predicates in a sentence) and surface position
(in a left-to-right order) by mapping each
pronoun into 5 levels: a) rank 1 for pronouns
with semantic role ARG0/agent of the main
predicate; b) rank 2 for pronouns with semantic role ARG1/patient of the main predicate; c)
rank 3 for pronouns with semantic role
ARG0/agent of other predicates; d) rank 4 for
pronouns with semantic role ARG1/patient of
other predicates; e) rank 5 for remaining pronouns. Furthermore, for those pronouns with
the same ranking level, they are ordered according to their surface positions in a left-toright order, motivated by previous research
on the centering theory (Grosz et al. 1995).
3) Detailed pronominal subcategory features.
Given a pronominal expression, its detailed
pronominal subcategory features, such as
whether it is a first person pronoun, second
person pronoun, third person pronoun, neuter
pronoun or others, are explored to enhance
the discriminative power of both the semantic
role features and the relative pronominal
ranking feature, considering the predominant
importance of pronouns in tracking the local
focus structure in discourse.
4.3
Comparison with Previous Work
As a representative in explicitly employing semantic role labeling in coreference resolution,
Ponzetto and Strube (2006) explored two semantic role features to capture the predicateargument structure information to benefit
coreference resolution: I_SEMROLE, the predicate-argument pairs of one referring expression,
and J_SEMROLE, the predicate-argument pairs
of another referring expression. Their experiments on the ACE 2003 corpus shows that,
while the two semantic role features much improve the performance of common noun resolution by 3.8 and 2.7 in F-measure on the BNEWS
and NWIRE domains respectively, they only
1
According to the centering theory, the backward-looking
center Cb is preferentially realized by a pronoun in the subject position in natural languages, such as English, and
people tend to plan ahead and minimize the number of
focus shifts.
991
slightly improve the performance of pronoun
resolution by 0.4 and 0.3 in F-measure on the
BNEWS and NWIRE domains respectively.
In comparison, this paper proposes various
kinds of centering-motivated semantic role features in attempt to better model the continuity or
shift in the local discourse focus by extending
the centering theory from the grammatical level
to the semantic level. For example, the
CAARG0MainVerb feature (as shown in Table
5) is designed to capture the semantic role of the
antecedent candidate in the main predicate in
modeling the discourse center, while, the ANPronounRanking feature (as shown in Table 5) is
designed to determinate the relative priority of
the pronominal anaphor in retaining the discourse center.
Although both this paper and Ponzetto and
Strube (2006) employs semantic role features,
their ways of deriving such features are much
different
due
to
different
driving
forces/motivations behind. As a result, their contributions on coreference resolution are different:
while the semantic role features in Ponzette and
Strube (2006) captures the predicate-argument
structure information and contributes much to
common noun resolution and their contribution
on pronoun resolution can be ignored, the centering-motivated semantic role features in this
paper contribute much in pronoun resolution.
This justifies our attempt to better model the
continuity or shift of the discourse focus in pronoun resolution by extending the centering theory from the grammatical level to the semantic
level and employing the centering-motivated
features in pronoun resolution..
5
Experimentation and Discussion
We have evaluated our approach of employing
the centering theory in pronoun resolution from
the semantic perspective on the ACE 2003 corpus.
5.1
Experimental Setting
The ACE 2003 corpus contains three domains:
newswire (NWIRE), newspaper (NPAPER), and
broadcast news (BNEWS). For each domain,
there exist two data sets, training and devtest,
which are used for training and testing respectively. Table 4 lists the pronoun distributions
with coreferential relationships in the training
data and the test data over pronominal subcategories and sentence distances. Table 4(a) shows
that third person pronouns occupy most and neu-
tral pronouns occupy second while Table 4(b)
shows that the antecedents of most pronouns
occur within the current sentence and the previous sentence, with a little exception in the test
data set of BNEWS.
NWIRE NPAPER BNEWS
Pronoun
Subcategory Train Test Train Test Train Test
First Person 263 103 283 120 455 258
Second Person 61
16 29 36 203 68
Third Person
618 179 919 263 736 158
Neuter
395 151 577 190 482 137
Reflexive
23
6
42 12 26 6
Other
0
0
2
0
2
3
(a) Distribution over pronominal subcategories
NWIRE NPAPER BNEWS
Train Test Train Test Train Test
0
890 254 1281 347 1149 295
1
447 149 529 197 729 188
2
0
27
0 24 0 41
>2
0
19
0 41 0 100
Total 1337 449 1810 609 1878 624
(b) Distribution over sentence distances
Table 4: Pronoun statistics on the ACE 2003 corpus
Distance
For preparation, all the documents in the corpus are preprocessed automatically using a pipeline of NLP components, including tokenization
and sentence segmentation, named entity recognition, part-of-speech tagging and noun phrase
chunking. Among them, named entity recognition, part-of-speech tagging and noun phrase
chunking apply the same Hidden Markov Model
(HMM)-based engine with error-driven learning
capability (Zhou and Su, 2000 & 2002). In particular for SRL, we use a state-of-the-art inhouse toolkit, which achieved the precision of
87.07% for ARG0 identification and the precision of 78.97% for ARG1 identification, for easy
integration. In addition, we use the SVM-light 2
toolkit with the radial basis kernel and default
learning parameters. Finally, we report the performance in terms of recall, precision, and Fmeasure, where precision measures the percentage of correctly-resolved pronouns (i.e. correctly
linked with any referring expression in the
coreferential chain), recall measures the coverage of correctly-resolved pronouns, and Fmeasure gives an overall figure on equal harmony between precision and recall. To see
whether an improvement is significant, we also
conduct significance testing using paired t-test.
In this paper, ‘>>>’, ‘>>’ and ‘>’ denote pvalues of an improvement smaller than 0.01, inbetween (0.01, 0,05] and bigger than 0.05,
2
http://svmlight.joachims.org/
992
which mean significantly better, moderately
better and slightly better, respectively.
5.2
Experimental Results
Table 5 details various centering-motivated features from the semantic perspective, which are
incorporated in our final system. For example,
the CAARG0MainVerb feature is designed to
capture the semantic role of the antecedent candidate in the main predicate in modeling the discourse center, while the ANPronounRanking
feature is designed to determinate the relative
priority of the pronominal anaphor in retaining
the discourse center. As the baseline, we duplicated the representative system with the same set
of 12 basic features, as described in Soon et al
(2001). Table 6 shows that our baseline system
achieves the state-of-the-art performance of 62.3,
65.3 and 59.0 in F-measure on the NWIRE,
NPAPER and BNEWS domains, respectively. It
also shows that the centering-motivated features
(from the semantic perspective) significantly
improve the F-measure by 3.6(>>>), 4.5(>>>)
and 7.7(>>>) on the NWIRE, NPAPER and
BNEWS domains, respectively. This justifies
our attempt to model the continuity or shift of
the discourse focus in pronoun resolution via
centering-motivated features from the semantic
perspective. For comparison, we also evaluate
the performance of our final system from the
grammatical perspective. This is done by replacing semantic roles with grammatical roles in
deriving centering-motivated features. Here, labeling of grammatical roles is achieved using a
state-of-the-art toolkit, as described in Buchholz
(1999). Table 6 shows that properly employing
the centering theory in pronoun resolution from
the grammatical perspective can also improve
the performance. However, the performance improvement of employing the centering theory
from the grammatical perspective is much lower,
compared with that from the semantic perspective. This validates our attempt of employing the
centering theory in pronoun resolution from the
semantic perspective instead of from the grammatical perspective. This also suggests the great
potential of applying the centering theory in
pronoun resolution since the centering theory is
a local coherence theory, which tells how subsequent utterances in a text link together.
Table 7 shows the contribution of the semantic role features and the relative pronominal
ranking feature in pronoun resolution when the
detailed pronominal subcategory features are
included:
Feature category
Feature
Remarks
1 if the semantic role of the antecedent candidate is
CAARG0
ARG0/agent; else 0
Semantic Role-based Fea1 if the antecedent candidate has the semantic role of
CAARG0MainVerb
tures
ARG0/agent for the main predicate of the sentence; else 0
1 if the anaphor and the antecedent candidate share the same
ANCASameTarget
predicate with regard to their semantic roles; else 0
Relative Pronominal RankWhether the pronominal anaphor is ranked highest among all
ANPronounRanking
the pronouns in the sentence
ing Feature
Whether the anaphor is a first person, second person, third
ANPronounType
Detailed Pronominal Subperson, neuter pronoun or others
category Features
Whether the antecedent candidate is a first person, second
CAPronounType
person, third person, neuter pronoun or others
Table 5: Centering-motivated features incorporated in our final system
(with AN indicating the anaphor and CA indicating the antecedent candidate)
System Variation
NWIRE
NPAPER
R% P%
F
R% P%
F
57.0 68.6 62.3 61.1 70.1 65.3
64.1 67.8 65.9 67.5 72.4 69.8
R%
49.0
59.9
Baseline System
Final System
(from the semantic perspective)
Final System
63.3 64 63.6 64.7 68.8 66.7 57.1
(from the grammatical perspective, for comparison)
Table 6: Contributions of centering-motivated features in pronoun resolution
BNEWS
P%
F
73.9 59.0
75.3 66.7
70.1
63.1
NWIRE
NPAPER
BNEWS
R% P%
F
R% P%
F
R% P%
F
Baseline System
57.0 68.6 62.3 61.1 70.1 65.3 49.0 73.9 59.0
+SR and DC
64.8 67.8 66.3 67.2 72.9 69.9 59.1 75.3 66.3
+PR and DC
61.5 65.4 63.4 64.9 72.1 68.3 57.4 73.5 64.5
+SR, PR and DC (Final System) 64.1 67.8 65.9 67.5 72.4 69.8 59.9 75.3 66.7
Table 7: Contribution of the semantic role features (SR) and the relative pronominal ranking feature (PR) in pronoun resolution when the detailed pronominal subcategory features are included
System Variation
1) The inclusion of the semantic role features
improve the performance by 4.0(>>>),
4.6(>>>) and 7.3(>>>) in F-measure on the
NWIRE, NPAPER and BNEWS domains, respectively. This suggests the impact of semantic role information in determining the
local discourse focus. Since pronouns preferentially occur in the subject position and tend
to refer to the main subject (Ehrlich 1980;
Brennan 1995; Walker et al. 1998; Cahn
1995; Gordon and Searce 1995; Kibble et al.
2001), this paper only applies semantic features related with the semantic role of
ARG0/agent, which is closely related with
the grammatical role of subject, with regard
to various predicates in a sentence. We have
also explored features related with other semantic roles. However, our preliminary experimentation shows that they do not improve
the performance, even for ARG1/patient, and
thus are not included in the final system. This
may be due to that other semantic roles are
993
not discriminative enough to make a difference in deciding the local discourse structure.
2) It is surprising to notice that further inclusion
of the relative pronominal ranking feature has
only slight impact (slight positive impact on
the BNEWS domain and slight negative impact on the NWIRE and NPAPER domains)
on the ACE 2003 corpus. This suggests that
most of information in the relative pronominal ranking feature has been covered by the
semantic role features. This is not surprising
since the semantic role of ARG0/agent,
which is explored to derive the semantic role
features, is also applied to decide the relative
pronominal ranking feature.
The inclusion of the relative pronominal
ranking feature improve the performance by
1.1(>>>), 3.0(>>>) and 5.5(>>>) in F-measure.
Our further evaluation reveals that the performance improvement difference among different
domains of the ACE 2003 corpus is due to the
distribution of pronouns’ antecedents occurring
over different sentence distances, as shown in
Final System
Baseline System
Pronoun
NWIRE NPAPER BNEWS
Subcategory
First Person
55.7
55.9
56.6
Second Person 54.6
60.4
44.0
Third Person
72.6
80.9
75.7
Neuter
41.5
50.4
50.2
Reflexive
60.0
85.7
70.0
Total
62.3
65.3
59.0
First Person
64.7
67.0
65.6
Second Person 78.6
70.0
51.9
Third Person
80.9
81.8
80.4
Neuter
48.3
53.0
58.3
Reflexive
71.4
66.7
80.0
Total
65.9
69.8
66.7
Table 8: Performance comparison of pronoun resolution in F-measure over pronoun subcategories
Table 8 shows the contribution of the centering-motivated features over different pronoun
subcategories. It shows that the centeringmotivated features contribute much to the resolution of the four major pronoun subcategories
(i.e. first person, second person, third person and
neuter) while its negative impact on the minor
pronoun subcategories (e.g. reflexive) can be
ignored due to their much less frequent occur994
rence in the corpus. In particular, the centeringmotivated features improve the performance on
the major three pronoun subcategories of third
person / neuter / first person, by
8.3(>>>)/6.8(>>>)/9.0(>>>),
0.9(>>)/
2.6
(>>>)/11.1(>>>) and 4.7(>>>)/8.1(>>>)/9.0
(>>>), on the NWIRE, NPAPER and BNEWS
domains of the ACE 2003 corpus, respectively.
Baseline
System
Distance NWIRE NPAPER BNEWS
<=0
61.6
64.5
68.7
<=1
60.4
67.5
60.0
<=2
62.9
67.4
63.7
Total
62.3
65.3
59.0
<=0
64.3
70.3
78.7
<=1
66.8
72.3
72.5
<=2
66.6
71.8
71.8
Total
65.9
69.8
66.7
Table 9: Performance comparison of pronoun resolution in F-measure over sentence distances
Final System
Table 4. This suggests the usefulness of the relative pronominal ranking feature in resolving
pronominal anaphors over longer distance. This
is consistent with our observation that, as the
percentage of pronominal anaphors referring to
more distant antecedents increase, its impact
turns gradually from negative to positive, when
further including the relative pronominal ranking
feature after the semantic role features. The reason that we include the detailed pronominal subcategory information is due to predominant
importance of pronouns in tracking the local
focus structure in discourse and that such detailed pronominal subcategory information is
discriminative in tracking different subcategories of pronouns. This suggests the usefulness of
considering the distribution of the local discourse focus over detailed pronominal subcategories. One interesting finding in our
preliminary experimentation is that the inclusion
of the detailed pronominal subcategory features
alone even harms the performance. This may be
due to the reason that the detailed pronominal
subcategory features do not have the discriminative power themselves and that the semantic role
features and the relative pronominal ranking feature provide an effective mechanism to explore
the role of such detailed pronominal subcategory
features in helping determine the local discourse
focus.
Table 9 shows the contribution of the centering-motivated features over different sentence
distances. It shows that the centering-motivated
features improve the performance of pronoun
resolution on different sentence distances of
0/1/2, by 2.7(>>>) / 5.8(>>>) / 10.0 (>>>),
6.4(>>>) / 4.8(>>>) / 12.5(>>>) and 3.7
(>>>)/4.4(>>>)/8.1(>>>), on the NWIRE,
NPAPER and BNEWS domains of the ACE
2003 corpus, respectively. This suggests that the
centering-motivated features are helpful for both
intra-sentential and inter-sentential pronoun
resolution.
6
Conclusion and Further Work
This paper extends the centering theory from the
grammatical level to the semantic level and
much improves the performance of pronoun
resolution via centering-motivated features from
the semantic perspective. This is mainly realized
by employing various semantic role features
with regard to various predicates in a sentence,
in attempt to model the continuity or shift of the
local discourse focus. Moreover, the relative
ranking feature of a pronoun among all the pronouns is explored to help determine the relative
priority of the pronominal anaphor in retaining
the local discourse focus. Evaluation on the
ACE 2003 corpus shows that both the centeringmotivated semantic role features and pronominal
ranking feature much improve the performance
of pronoun resolution, especially when the detailed pronominal subcategory features of both
the anaphor and the antecedent candidate are
included. It is not surprising due to the predomi-
nance of pronouns in tracking the local discourse
structure in a text.
To our best knowledge, this paper is the first
research which successfully applies the centering-motivated features in pronoun resolution
from the semantic perspective.
For future work, we will explore more kinds
of semantic information and structured syntactic
information in pronoun resolution. In particular,
we will further employ the centering theory in
pronoun resolution from both grammatical and
semantic perspectives on more corpora.
Acknowledgement
This research is supported by Project 60673041
under the National Natural Science Foundation
of China, project 2006AA01Z147 under the
“863” National High-Tech Research and Development of China, project 200802850006 under
the National Research Foundation for the Doctoral Program of Higher Education of China,
project 08KJD520010 under the Natural Science
Foundation of the Jiangsu Higher Education Institutions of China.
References
C. Aone and W.W. Bennett. 1995. Evaluating automated and manual acquisition of anaphora resolution strategies. ACL’1995:122-129.
S.E. Brennan, M.W. Friedman, and C.J. Pollard. 1987.
A centering approach to pronoun. ACL’1987: 290297.
S.E. Brennan 1995. Centering attention in discourse.
Language and Cognitive Process, 10/2: 137-67.
S. Buchholz, J. Veenstra and W. Daelemans. 1999.
Cascaded Grammatical Relation Assignment.
EMNLP-VLC’1999: 239-246
S. Cote. 1983. Ranking and forward-looking centers.
In Proceedings of the Workshop on the Centering
Theory in Naturally-Occurring Discourse. 1993.
D.M. Carter. 1987. Interpreting Anaphors in Natural
Language Texts. Ellis Horwood, Chichester, UK.
J. Carbonell and R. Brown. 1988. Anaphora resolution: a multi-strategy approach. COLING’1988:
96-101.
B.J. Grosz, A.K. JoShi and S. Weinstein. 1995. Centering: a framework for modeling the local coherence of discourse. Computational Linguistics,
21(2):203-225.
P.C. Gordon, B.J. Grosz and L.A. Gilliom. 1993.
Pronouns, names and the centering of attention in
discourse. Cognitive Science.1993.17(3):311-348
P.C. Gordon and K. A. Searce. 1995. Pronominalization and discourse coherence, discourse structure
and pronoun interpretation. Memory and Cognition. 1995.
995
S. Harabagiu and S. Maiorano. 2000. Multiligual
coreference resolution. ANLP-NAACL 2000:142149.
A.K. Joshi and S. Weinstein. 1981. Control of inference: Role of some aspects of discourse structurecentering. IJCAI’1981:385-387
R. Kibble. 2001. A Reformulation of Rule 2 of Centering. Computational Linguistics, 2001,27(4):
579-587
M. Kameyama. 1986. Aproperty-sharing constraint in
centering. ACL 1986:200-206
M. Kameyama. 1988. Japanese zero pronominal
binding, where syntax and discourse meet. In Proceeding of the Second International Workshop on
Japanese Syntax. 1988.
R. Mitkov. 1998. Robust pronoun resolution with
limited knowledge. COLING-ACL’1998:869-875.
A. Moschitti and S. Quarteroni. 2008. Kernels on
linguistics structures for answer entraction.
ACL’08:113-116
V. Ng and C. Cardie. 2002. Improving machine
learning approaches to coreference resolution.
ACL’2002: 104-111
V. Ng. 2007. Semantic Class Induction and Coreference Resolution. ACL’2007 536-543.
M. Palmer, D. Gildea and P. Kingsbury. 2005. The
proposition bank: A corpus annotated with semantic roles. Computational Linguistics, 31(1):71-106.
S. Pradhan, K. Hacioglu, V. Krugler, W. Ward, J.H.
Martin, and D. Jurafsky. 2005. Support vector
learning for semantic argument classification. Machine Learning, 2005.60:11-39.
S. P. Ponzetto and M. Strube. 2006. Semantic Role
Labeling
for
Coreference
Resolution.
EMNLP’2006 143-146.
E. F. Prince and M. A. Walker. 1995. A bilateral apporach to givenness: a hearer-status algorithm and
a centering algorithm. In Proceedings of 4th International Pragmatics Conference.
E. Rich and S. LuperFoy. 1988. An architecture for
anaphora resolution. In Proceedings of the 2nd
Conference on Applied Natural Language Processing. ANLP’1988: 18-24.
W.M. Soon, H.T. Ng and D. Lim. 2001. A machine
learning approach to coreference resolution of
noun phrase. Computational Linguistics, 2001,
27(4):521-544.
D. Shen and M. Lapata. 2007. Using semantic roles
to improve question answering. EMNLP-CoNIL
2007:12-21
C. Sidner. 1979. Toward a computation of intrasentential coreference. Technical Report TR-537,MIT.
Artificial Intelligence Laboratory.
C. Sidner. 1981. Focusing for interpretation of pronouns. Computational Linguistics,1981.7:217-231
J. Tetreault. 1999. Analysis of syntax-based pronoun
resolution methods. ACL 1999:602-605
J. Tetreault. 2001. A corpus-based evaluation of centering and pronoun resolution. Computational Linguistics. 2001. 27(4):507-520.
M. Walker, A. K. Joshi and E. Prince. 1998. Centering in naturally occurring discourse: An overview.
Clarendon Press:1-28
X.F. Yang，J. Su，G.D. Zhou and C.L. Tan. 2004.
Improving pronoun resolution by incorporating
coreferential
information
of
candidates.
ACL’2004:127-134.
X.F. Yang， J. Su and C.L. Tan. 2005. Improving
Pronoun Resolution Using Statistics - Based Semantic Compatibility Information. ACL’2005:165
-172.
X.F. Yang and J. Su. 2007. Coreference Resolution
Using Semantic Relatedness Information from
Automatically Discovered Patterns. ACL’2007:
528-535.
G.D. Zhou and J. Su. 2004. A high- performance
coreference resolution system using a multi- agent
strategy. COLING’ 2004:522- 528.
Y. Ziv and B.J. Grosz. 1994. Right dislocation and
attentional state. Israel Association of Theoretical
Linguistics Meetings’1994. 184-199.
996